Deep Learning-Based Models For Real-World Time Series Classification And Forecasting

Time series classification and forecasting are significant research areas that researchers often neglected.The abandonment of the mentioned above problems is because the time component causes time series problems to be too hard to handle.However,using the propounded classification and fore-casting models of this thesis can help address the above issues.It can also play a constructive part in the analysis of real-world time series in distinct application domains.Deep Learning schemes are becoming progressively crowd-pleasing in solving real-world time series problems,specifically classification and forecasting related problems.Datasets for time series are often available online free while others are not open source.However,due to the widespread use of time series,the need arises for accurate classification and forecasting.This thesis combined the two problems as one problem named Deep Learning-Based Models for Real-World Time Series Classification and Forecasting.They utilise the entire instances of datasets and,at the same time,defeated existing baseline schemes.Nonetheless,our deep learning networks are severely affeected by the problem of overfitting.For the past decade,we have noticed the utilisation of linear schemes in various disciplines.Not too long ago,a subject called deep learning presented opportunities to input dataset into an algorithm minus any exhaustive car-pentered attribute engineering approaches.This work proposed an architecture that captures attributes in a self-supervised(i.e.,unsupervised)manner via the Contextual Convolutional Neural Networks(CCNN)and Contextual Long Short-Term Memory(CLSTM)arms to address the time series classification efficiently.Using the idea of contextual features in the two feature extraction arms,the capturing of attributes is efficient.The final stage is a stand-alone Multilayer Perceptron(MLP)block,which executes the classification.Altering the number of neurons and the placement of dropouts are the two measures taken to fix overfitting.Lastly,assessments regarding the University of California Riverside(UCR)data set unveils our scheme's excellence.Our second work proffer solution or advice to an actual investment problem named days on the market.It is no secret that the Days on Market(DOM)feature has been a very instrumental statistic for the real estate industry because of its role in assessing homes in a real estate firm.For the second work,we propounded an innovative hybrid of deep learning scheme,purposefully designed to address the DOM prediction problem adequately.Our proposed algorithm captures attributes through the CNN-based Attention(CNNA),plus Bidirectional LSTM(BLSTM)modules.Besides,we present confidence intervals for the dataset's four attributes by either percentile boot-strap confidence interval(CI)or percentile bias-corrected accelerated(BCa)bootstrap CI decided by an estimated distribution for a feature.Finally,we appraise our algorithm through experimentation with a sure-enough dataset of a notable real estate establishment in Shanghai,China.Our experimental results reveal the superbness of our projected scheme in solving the DOM prediction problem.However,in the recent past,modelling and predicting stock prices have been a severe challenge for both the business community and the research domain due to noise and the non-stationary character of data instances.In the third work,we projected an innovative end-to-end algorithm named SMF to fix the stock price forecasting problem.The attributes are captured by both AB-CNN and CB-LSTM arms.Nonetheless,both LSTM and Attention mechanism are incorporated in the structure to capture long-range,including extraordinarily long-term stock information,respectively.We implemented our propounded scheme on a popular stock code-named "JNJ",a famous NYSE stock market member.Experimental results divulge that our pro-posed algorithm outperforms stand-alone deep learning schemes,statistical techniques,as well as machine learning approaches on all accounts.In con-clusion,we also appraise our stock dataset to point out short-term trading opportunities via the use of indicators from Technical Analysis(TA).Namely,Moving Average(MA),Moving Average Convergence-Divergence(MACD)curve,MACD histogram,Relative Strength Index(RSI),etc.In the above three experiments,we investigated distinct hybrid deep learning schemes in three different disciplines.The above three fields/problems are as follows,time series classification,forecasting days on market feature for real estate firms,and predicting stock prices in a stock market.However,the three experiments' results unveil the unique nature of both stand-alone and hybrid deep learning schemes in solving real-world time series problems.